Novel Role Filler Generalization for Recurrent Neural Networks Using Working Memory-Based Indirection

Abstract

Humans encounter and adapt to novel situations every day. However, adaptation is not a trivial task to accomplish. In the field of machine learning, the statistical underpinnings of established deep learning architectures make it difficult for these architectures to handle certain types of novel situations. Previous research demonstrates how computational models could better handle novel situations through indirection, an idea inspired by the interaction between two regions of the human brain: the prefrontal cortex and the basal ganglia. This thesis demonstrates that combining the indirection model with deep learning methods outperforms current architectures.